iFlow Search MCP Server

iflow-search-js

JavaScript / TypeScript integrations for iFlow Search (心流搜索) — a search API that provides web search, image search, and web page fetching with AI-friendly structured output.

• Product: https://platform.iflow.cn/
• API docs: https://platform.iflow.cn/docs/
• Operator: 杭州星辰千寻科技有限公司

This monorepo contains:

• A framework-agnostic core SDK
• A LangChain JS tool adapter (also usable from LangGraph)
• A LangGraph agent example
• An MCP stdio server for Hermes / Claude Code / Claude Desktop / other MCP clients

Package / usage matrix

Why there is no @iflow-ai/search-langgraph

LangGraph consumes LangChain tools directly. A separate search-langgraph package would be a thin re-export with no added value — use @iflow-ai/search-langchain for both. See examples/langgraph-agent for a working createReactAgent wiring.

Current status

• ✅ @iflow-ai/search-core — implemented, framework-agnostic client
• ✅ @iflow-ai/search-langchain — implemented, three tools (iflow_web_search, iflow_image_search, iflow_web_fetch)
• ✅ examples/langgraph-agent — implemented, ReAct agent end-to-end smoke validated against real iFlow API
• ❌ no separate @iflow-ai/search-langgraph package (intentional — see above)
• ✅ @iflow-ai/search-mcp — implemented, stdio MCP server with three tools mirroring the LangChain adapter; optional MCP-host attribution via IFLOW_MCP_CLIENT / IFLOW_MCP_CLIENT_VERSION
• ✅ plugins/iflow-search — Claude Code Plugin manifest (metadata-only, runs npx -y @iflow-ai/search-mcp); awaiting submission to the Anthropic Plugin Directory. See plugins/iflow-search/README.md.

Workspace development

pnpm install
pnpm -r run typecheck
pnpm -r run build
pnpm -r run test

The workspace pins @langchain/core to ^1.1.44 via pnpm-workspace.yaml overrides: to keep LangGraph's runtime and LangChain's tool types on a single major. Removing this override re-introduces ToolMessage cross-version serialization bugs in LangGraph agent loops; rerun the LangGraph agent tests before you touch it.

Basic usage — core SDK

import { createIFlowSearchClient } from "@iflow-ai/search-core";

const client = createIFlowSearchClient({
  apiKey: process.env.IFLOW_API_KEY!,
  source: "core",
  integrationName: "my-app",
  integrationVersion: "1.0.0",
});

const result = await client.webSearch({ query: "flash attention", count: 5 });
if (!result.ok) {
  console.error(result.error.code, result.error.message);
} else {
  for (const r of result.data.results) console.log(r.title, r.url);
}

See packages/search-core/README.md for the full API surface.

Basic usage — LangChain

import { createIFlowSearchTools } from "@iflow-ai/search-langchain";

const tools = createIFlowSearchTools({
  apiKey: process.env.IFLOW_API_KEY!,
});

// Hand `tools` to any LangChain JS agent that supports `bindTools`.

See packages/search-langchain/README.md for tool-by-tool docs.

Basic usage — LangGraph

LangGraph does not need a separate iFlow package. Pass the same @iflow-ai/search-langchain tools into createReactAgent:

import { createReactAgent, ToolNode } from "@langchain/langgraph/prebuilt";
import { createIFlowSearchTools } from "@iflow-ai/search-langchain";

const tools = createIFlowSearchTools({ apiKey: process.env.IFLOW_API_KEY! });
const agent = createReactAgent({ llm, tools: new ToolNode(tools) });

The working end-to-end example lives at examples/langgraph-agent — bring your own tool-calling LLM (DeepSeek, OpenAI, Anthropic, …).

Attribution headers

Every request to iFlow goes through @iflow-ai/search-core and carries:

IFlow-Source:              <runtime>
IFlow-Integration:         <package-name>
IFlow-Integration-Version: <package-version>
User-Agent:                <package-name>/<package-version>

Currently emitted values:

@iflow-ai/search-mcp additionally emits IFlow-MCP-Client and IFlow-MCP-Client-Version when the MCP host declares itself via IFLOW_MCP_CLIENT / IFLOW_MCP_CLIENT_VERSION environment variables (e.g. hermes, claude-code, claude-desktop). Absence of these headers is meaningful — there is no unknown placeholder, so hosts that opt out remain indistinguishable from hosts that have not adopted the convention.

LangGraph traffic shows up as IFlow-Source: langchain because it consumes the same LangChain adapter — there is no separate langgraph source ID for this reason.

Security / key handling

• Never commit real API keys. Use environment variables (IFLOW_API_KEY, plus your LLM provider key for agent examples).
• Don't write keys into .env files that get committed, into package.json, or into test fixtures.
• The unit tests in this repo use fake keys and a mocked fetch — they never touch the real iFlow API.
• Real-API smoke tests are opt-in. The committed smoke scripts (e.g. packages/search-langchain/scripts/smoke-direct.mjs) read IFLOW_API_KEY from the environment at runtime and never persist it.

Roadmap

• P0 ✅ @iflow-ai/search-core — framework-agnostic SDK
• P1 ✅ @iflow-ai/search-langchain — LangChain JS tool adapter
• P2 ✅ examples/langgraph-agent — LangGraph ReAct agent example
• P3 ✅ @iflow-ai/search-mcp — MCP server for Hermes / Claude Desktop / other MCP clients
• P4 🟡 docs and examples polish — broader recipes, additional LLM providers
• P5 optional — refactor the OpenClaw community iFlow plugin to reuse @iflow-ai/search-core

Maintenance and roadmap

Long-form docs for maintainers and contributors live under docs/:

• docs/package-strategy.md — which npm packages we publish, which we explicitly will not, and the decision rubric for new ones.
• docs/integration-roadmap.md — how each framework (LangChain, LangGraph, OpenClaw, Hermes, Claude Code, iFlow CLI, Open WebUI, Coze, CrewAI) reaches iFlow Search, with phased execution order.
• docs/release-policy.md — versioning, next vs latest, the per-release checklist, and the manual publish commands.
• docs/mcp-design.md — design for the planned @iflow-ai/search-mcp server (P3): package decision, transport scope, tool schema, attribution headers, MCP client config example, and test strategy.

License

MIT — see LICENSE.